Non-sensing management method and system for costume assets based on time sequence video stream

Abstract

The present application relates to the technical field of digital wardrobe asset management, and discloses a clothing asset non-sensing management method and system based on time sequence video streams; the clothing asset non-sensing management method is applied to an asset non-sensing management device and specifically comprises the following steps: S11: collecting time sequence video streams of a user performing a taking and placing operation process on a wardrobe through a single static camera deployed in the wardrobe, and performing real-time frame-by-frame inference on each frame of image in the collected time sequence video streams. The present application continuously observes the natural taking and placing behavior of the user through the single static camera, locks the interaction relationship between the hand and the clothes in real time under the premise that the user is completely non-sensing, automatically distinguishes new clothes and historical clothes by using appearance feature extraction and database similarity comparison, and accurately determines the two types of operation gestures of taking out and putting in according to a motion recognition model, so as to drive a state machine to complete the real-time flipping of the in-stock state bit.

Description

Technical Field

[0002] This invention relates to the technical field of digital wardrobe asset management, and more specifically, to a method and system for seamless management of clothing assets based on time-series video streams. Background Technology

[0003] With the rapid development of artificial intelligence and the Internet of Things (IoT) technologies, smart wardrobes and digital wardrobe management systems have gradually become research hotspots in the fields of smart homes and fashion technology. In clothing image recognition and retrieval, deep learning-based clothing target detection models can identify individual clothing items from user-uploaded images and cut them out as input for feature extraction models. Through deep feature extraction and similarity matching, automatic identification and classification of clothing attributes such as category, color, and style are achieved. The public availability of large-scale clothing datasets provides a training foundation for tasks such as clothing detection, recognition, retrieval, and verification. At the smart wardrobe system level, existing solutions use semantic segmentation models to identify and classify user-uploaded clothing images, achieving digital inventory management. Some solutions combine RFID tags with sensor networks, attaching electronic tags to each garment to achieve clothing identification and location tracking. Other solutions use reed switches or mechanical switching of location tags to determine the usage status of clothing trays. In terms of outfit recommendations, intelligent matching engines based on large language models or multimodal models can automatically generate matching suggestions based on the structured attribute information of clothing.

[0004] Although smart wardrobes and digital wardrobe management systems have made significant progress, they still have the following two shortcomings:

[0005] First, the process of digitizing clothing assets is extremely cumbersome and heavily relies on manual operation by users. Whether it is an image recognition-based solution or an RFID-based solution, users need to actively cooperate. The former requires users to take photos of each piece of clothing, upload them, manually cut out the background and label them with category tags. The latter requires users to manually attach or sew RFID tags to each newly purchased piece of clothing and complete the initial binding. This process is not only time-consuming and laborious, but more seriously, the system cannot automatically detect when users have purchased new clothing or when they have hung the new clothing in the closet, resulting in a natural time lag and operational gap in the entry of new assets.

[0006] Secondly, the in-stock / out-of-stock status of clothing is severely disconnected from reality, making it impossible to perceive users' daily clothing taking and putting away behavior in real time and seamlessly. When a user takes a piece of clothing out of the closet and wears it out, the status record in the system database is not automatically updated. The clothing is still marked as being in the closet in the digital world. This disconnect directly causes the cloud-based outfit recommendation model to often recommend clothes that have already been worn out by the user but are not actually in the closet when making generative recommendations, which greatly weakens the actual usability and user experience of intelligent outfit recommendations.

[0007] In summary, existing technologies all require user-initiated actions as a necessary prerequisite for data entry and status updates, rather than using the user's natural picking and putting behavior as a trigger signal, thus failing to achieve truly seamless asset management. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] The purpose of this invention is to provide a method and system for seamless management of clothing assets based on time-series video streams. It automatically distinguishes between new and historical clothing by extracting appearance features and comparing similarity with the database, and accurately determines two types of operation gestures, namely taking out and putting in, based on the action recognition model. This drives the state machine to complete the real-time flipping of the inventory status, thus eliminating the system's dependence on user-initiated operations.

[0010] (II) Technical Solution

[0011] This invention is implemented as follows: a method for seamless management of apparel assets based on time-series video streams, applied to seamless asset management devices, specifically including the following steps:

[0012] S11: A single static camera deployed in the wardrobe captures a time-series video stream of the user's actions of taking and putting clothes in the wardrobe. The system performs real-time frame-by-frame reasoning on each frame of the captured time-series video stream, tracks the motion trajectory between the user's hand area and the target area of ​​the clothing, and the pixel-level appearance feature sequence of the clothing target in the time series. Based on the motion trajectory, the system initially determines the behavior category of the clothing target involved in the current operation in the time series.

[0013] S12: Based on the extracted appearance feature sequence of the clothing target in time, the appearance feature sequence is mapped into a clothing feature vector with a preset dimension through a deep feature extraction network. The clothing feature vector is then compared with the feature vectors of all clothing assets stored in the cloud database to obtain a similarity measurement value sequence.

[0014] S13: Based on the similarity measurement value sequence and the preset overlap threshold, the asset feature attributes of the clothing target are obtained, the identity of the current clothing target is locked through the asset feature attributes, and a globally unique physical asset identifier is automatically generated for it in the cloud database.

[0015] S14: After the identity of the current clothing target is locked, based on the relative spatial position relationship and movement trajectory between the user's hand and the clothing target obtained by tracking, the user's current operation gesture is classified by the deployed temporal behavior recognition model. According to the judgment result, the in-stock status bit corresponding to the clothing target in the database is automatically updated. Each flip of the in-stock status bit is accompanied by a timestamp record, forming a complete temporal log of the status change of the clothing asset.

[0016] S15: Receive an outfit recommendation request initiated by the outfit recommendation module deployed in the cloud. The outfit recommendation module uses the list of physical asset identifiers as the candidate recommendation SKU pool, performs outfit collaborative generation reasoning based on structured attribute information, and outputs an outfit recommendation scheme that only contains clothing currently in stock.

[0017] Furthermore, in S11, a single static camera deployed inside the wardrobe captures a time-series video stream of the user's actions of taking and placing items in the wardrobe. Real-time frame-by-frame inference is performed on each frame of the captured time-series video stream, including:

[0018] A single static camera is fixedly deployed in the top corner of the wardrobe or in an external position facing the wardrobe door opening. Its optical axis is directed towards the clothing retrieval and placement area of ​​the wardrobe. The camera continuously captures a time-series video stream of the complete operation process of the user opening the wardrobe door, taking out and placing clothes, and closing the wardrobe door at a preset frame rate of 15 to 30 frames per second. The preset frame rate is dynamically and adaptively adjusted according to the current computing load of the end device. When the load is lower than the first threshold, the first frame rate value is used. When the load is higher than the first threshold but lower than the second threshold, the second frame rate value is used. When the load is higher than the second threshold, a frame skipping strategy is triggered, and only key frames are captured and inferred. Non-key frames are estimated by interpolation.

[0019] Based on the deployed target detection and target tracking algorithms, real-time frame-by-frame inference is performed on each frame of the acquired time-series video stream. The target detection algorithm performs forward inference on each frame of the image using a single-stage detection framework, and outputs the bounding box coordinates and category confidence of the user's hand and clothing targets.

[0020] The target tracking algorithm predicts the target's position in the next frame based on the intersection-union ratio (IU) matching between the detection box in the current frame and the tracked target in the historical frames, thus maintaining consistent tracking of the same target across consecutive frames.

[0021] Furthermore, the tracking of the motion trajectory between the user's hand area and the target clothing area, as well as the pixel-level appearance feature sequence of the clothing target over time, includes:

[0022] After the target detection algorithm outputs the bounding box coordinates and category confidence scores of the user's hand and clothing targets, redundant boxes are removed by non-maximum suppression and cross-frame identity association. The two-dimensional bounding box sequence of the user's hand region and clothing target region on the image plane is output in real time. Based on the two-dimensional bounding box sequence, the relative displacement, overlap area ratio and azimuth change between the hand center point and the clothing bounding box are calculated. The motion trajectory parameter sequence of the hand and clothing is generated based on the pixel displacement vector of the bounding box center point between consecutive frames.

[0023] The target tracking algorithm maintains consistent tracking of the same target across frames. For clothing targets with locked temporary tracking identifiers, pixel-by-pixel spatial sampling is performed within the pixel area covered by their bounding box to extract the pixel-level appearance feature sequence of the clothing target in the continuous time series.

[0024] The appearance feature sequence includes at least the color distribution features, texture features, edge contour features, and local key point features of the clothing. The color distribution features are obtained by converting the pixels within the bounding box from the RGB color space to the HSV color space and then calculating the pixel frequency distribution of each hue range. The texture features are obtained by performing local binary pattern encoding on the pixels within the bounding box and calculating their histogram distribution. The edge contour features are obtained by extracting the edge response map using the Canny edge detection operator and calculating the edge direction gradient histogram. The local key point features are obtained by detecting local key points within the bounding box region and extracting their 128-dimensional descriptors using the scale-invariant feature transform algorithm.

[0025] Furthermore, in S12, the similarity between the clothing feature vector and the feature vectors of all clothing assets stored in the cloud database is calculated one by one to obtain a sequence of similarity metrics, including:

[0026] The similarity of the current clothing feature vector with the feature vectors of all existing clothing assets stored in the cloud database is calculated one by one. The triple loss function is used to constrain the features of the same clothing from different perspectives to be close to each other in the feature space and the features of different clothing to be far apart. The similarity calculation adopts the cosine similarity measurement method, and the cosine value of the angle between the current feature vector and each existing feature vector is calculated as the similarity measurement value. After calculation, a similarity measurement value sequence containing the corresponding similarity values ​​of all existing clothing assets is obtained.

[0027] The maximum value in the similarity metric sequence is compared with a preset overlap threshold, which is dynamically and adaptively adjusted according to the total number of existing clothes in the database.

[0028] If the maximum value is lower than the overlap threshold, the current clothing target is determined to be a new clothing asset appearing for the first time, triggering the subsequent new asset entry process; if the maximum value is higher than or equal to the overlap threshold, the current clothing target is determined to be a historical clothing asset that has been recorded in the database, and the existing clothing asset corresponding to the maximum value is taken as a candidate matching asset, while the difference between the maximum value and the second largest value is recorded as a matching confidence index.

[0029] Furthermore, in S13, the identity of the current clothing target is locked through asset characteristic attributes, and a globally unique physical asset identifier is automatically generated for it in the cloud database, including:

[0030] If the current clothing target is determined to be a new clothing asset, then based on the extracted pixel-level appearance feature sequence, the structured attribute information of the clothing is automatically identified and extracted through the structured attribute recognition network deployed on the edge. The clothing feature vector is associated and bound with the structured attribute information, and a globally unique physical asset identifier is automatically generated for it in the cloud database, thus completing the seamless entry of the new clothing asset.

[0031] If the current clothing target is determined to be a historical clothing asset, then the current clothing target is redirected and associated with the existing physical asset identifier in the database through the physical asset identifier of the candidate matching asset, and all historical attribute information and status information of the existing clothing are reused.

[0032] Furthermore, in S14, based on the tracked relative spatial relationship and motion trajectory between the user's hand and the clothing target, the deployed temporal behavior recognition model classifies the user's current gesture, including:

[0033] Based on the relative spatial position relationship and motion trajectory, the relative displacement, overlapping area ratio, azimuth angle change, and hand displacement vector, velocity vector, and acceleration vector between the hand center point and the clothing bounding box are used as input features of the temporal behavior recognition model. The temporal video stream is uniformly divided into multiple video segments in the time dimension. Spatial features are extracted for each segment through a two-dimensional convolutional neural network. Then, the spatiotemporal features of multiple segments are aggregated through a temporal fusion layer. Finally, the probability distribution of the operation gesture belonging to the take-out or put-in category is output through a fully connected classification layer. The result is determined with the help of the trend of spatial position change before and after the hand operation and the trend of clothing bounding box size change.

[0034] Based on the determination result of the time-series behavior recognition model, the in-stock status position corresponding to the clothing target in the database is automatically updated. If it is determined that the current operation gesture belongs to the putting action category, indicating that the user is hanging the clothing back or putting it into the closet, the in-stock status position is automatically flipped to the first status value representing in-stock.

[0035] If the current gesture is determined to be a take-out action, indicating that the user is taking clothes out of the closet to wear, the in-stock status will be automatically flipped to the second status value representing out-of-stock.

[0036] Each flip of the in-stock status bit is accompanied by the current system timestamp and the corresponding video frame number record, which are appended to the status change time sequence log of the clothing asset in chronological order to form a status switching history record.

[0037] Furthermore, in S15, an outfit recommendation module deployed in the cloud is received, which generates an outfit scheme request. This module uses a list of physical asset identifiers as its candidate recommendation SKU pool, including:

[0038] The cloud-deployed outfit recommendation module responds to outfit scheme generation requests triggered by users or automatically triggered by the system according to a preset time strategy, and sends an instruction to the cloud database to query clothing assets in stock. The query instruction includes at least all clothing assets whose filtering condition is that the stock status bit value is a first status value representing the stock status.

[0039] The preset time strategy includes timed triggering at a fixed time every day, delayed triggering after the user leaves the wardrobe space, and conditional triggering triggered by changes in external environmental information.

[0040] The external environment information includes at least the weather information, temperature information, humidity information, and ultraviolet intensity information for the day. The query instruction also includes filtering conditions for one or more fields of color category, style category, and material category in the structured attribute information, which are automatically generated by the outfit recommendation module based on the external environment information and the user's historical outfit preference data.

[0041] Furthermore, a query command for in-stock clothing assets is sent to the cloud database, including:

[0042] In response to a query command, the cloud database returns a list of physical asset identifiers and corresponding structured attribute information for all clothing assets marked as in stock at the current moment, based on the system timestamp when the cloud database receives the query command.

[0043] The outfit recommendation module uses the list of physical asset identifiers returned by the query as a candidate recommendation SKU pool. Based on the preset outfit matching rule library, it determines the feasibility of pairing or multi-item combinations of candidate SKUs. The matching rule library includes at least color matching rules, style matching rules, material matching rules, and season matching rules. It also outputs one or more outfit recommendation schemes by sorting the feasible combinations according to the matching coordination score through collaborative filtering algorithm or sequence generation model based on attention mechanism.

[0044] All clothing assets marked as out of stock are automatically excluded from the candidate recommended SKU pool to ensure that the recommendation results are consistent with the actual physical asset status in stock in real time.

[0045] Compared with existing technologies, the apparel asset seamless management method and system based on time-series video streams provided by this invention have the following beneficial effects:

[0046] 1. By continuously observing the user's natural picking and placing behavior through a single static camera, the system uses edge target detection and tracking algorithms to lock the interaction between the hand and the clothing in real time without the user's awareness. It automatically distinguishes between new and old clothing by extracting appearance features and comparing similarity with the database. Based on the action recognition model, it accurately determines the two types of operation gestures of taking out and putting in, and drives the state machine to complete the real-time flipping of the inventory status. This eliminates the system's dependence on the user's active operation and integrates the asset management process into the user's daily life habits.

[0047] 2. By leveraging a lightweight convolutional neural network in collaboration with a segmented temporal behavior recognition model, stable real-time inference performance can be maintained even under limited computing resources. Feature vector retrieval and state filtering from the cloud database ensure that the outfit recommendation engine generates outfits based solely on the clothing assets currently in the database, maintaining consistency between the recommendation results and physical reality. Furthermore, the system automatically maintains a time-series log of state changes, including timestamps, using the unique physical asset identifier of each garment as its core. This provides a structured foundation for subsequent data value-added services such as wear frequency analysis and outfit effect tracking. This not only lowers the deployment threshold and usage cost of smart wardrobes but also lays the foundation for personalized outfit recommendations based on real-world data, enhancing user trust and satisfaction with smart outfit services.

[0048] A seamless apparel asset management system based on time-series video streams is used to execute the aforementioned seamless apparel asset management method. The management system includes:

[0049] The video acquisition module is used to continuously acquire the time-series video stream during the user's operation of taking and putting things in the wardrobe and transmit it to the terminal.

[0050] The edge processing module is used to perform real-time target detection and tracking on the time-series video stream, extract the pixel-level appearance feature sequence of clothing targets and map it into clothing feature vectors, identify the structured attribute information of clothing, and determine the category of taking out or putting in based on the user's current operation gesture.

[0051] The cloud database module is used to store the physical asset identifier, feature vector, structured attribute information and inventory status bits of all clothing assets, and maintain the status change time sequence log for each garment.

[0052] The deduplication and association module is used to compare the feature vector of the clothing with the existing feature vector in the database, determine whether the current clothing is a new asset or a historical asset, and trigger the entry of new assets or association with existing identifiers accordingly.

[0053] The status management module is used to flip the in-stock status bit of the corresponding clothing to in-stock or out-of-stock status based on the operation gesture judgment result.

[0054] The outfit recommendation module is used to respond to outfit generation requests by querying all clothing assets in the database and using the query results as a candidate recommendation pool to generate outfit recommendation schemes.

[0055] Specifically, the end-side processing module includes:

[0056] The target detection and tracking unit is used to perform target detection in each frame of the image, maintain consistent tracking of the same clothing target across consecutive frames, and output the motion trajectory of the hand and clothing.

[0057] The feature extraction unit is used to extract the pixel-level appearance feature sequence of the clothing target in time, map the appearance feature sequence into clothing feature vectors, and identify the structured attribute information of the clothing.

[0058] The behavior recognition unit is used to determine the category of taking out or putting in the user's operation gesture based on the motion trajectory. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the non-intrusive management method for apparel assets based on time-series video streams proposed in this invention.

[0060] Figure 2 This is a flowchart illustrating the process of initiating a query command for in-stock clothing assets to a cloud database in the non-intrusive management method for clothing assets based on time-series video streams proposed in this invention.

[0061] Figure 3 This is a schematic diagram of the structure of the apparel asset seamless management system based on time-series video streams proposed in this invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0063] The implementation of the present invention will be described in detail below with reference to specific embodiments.

[0064] In the accompanying drawings of this embodiment, the same or similar reference numerals correspond to the same or similar components. In the description of this invention, it should be understood that if terms such as up, down, left, right, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms describing positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting this invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0065] Reference Figure 1-2 As shown, in this embodiment, the apparel asset contactless management method based on time-series video streams is applied to an asset contactless management device, specifically including the following steps:

[0066] S11: A single static camera deployed inside the wardrobe captures a time-series video stream of the user's actions of taking and placing clothes in the wardrobe. Each frame of the captured video stream is analyzed in real-time, frame-by-frame, tracking the motion trajectory between the user's hand and the target clothing area, as well as the pixel-level appearance feature sequence of the clothing in the time series. Based on the motion trajectory, the temporal behavior category of the clothing involved in the current operation is initially determined. The single static camera is fixedly deployed in the top corner of the wardrobe or in an external position facing the wardrobe door opening, with its optical axis pointing towards the clothing taking and placing operation area. This ensures that the clothing target is within the effective field of view of the camera when the user performs the taking and placing operation, avoiding target loss due to obstruction of view. The camera continuously captures a time-series video stream of the complete operation process from opening the wardrobe door, taking and placing clothes to closing the wardrobe door at a preset frame rate of 15 to 30 frames per second, thus fully capturing the visual information throughout the entire cycle from the start to the end of the operation, providing sufficient temporal context for subsequent behavior recognition.

[0067] Specifically, the preset frame rate is dynamically and adaptively adjusted according to the current computing load of the edge device. When the load is below the first threshold, the first frame rate value is used; when the load is above the first threshold but below the second threshold, the second frame rate value is used; when the load is above the second threshold, a frame skipping strategy is triggered, and only key frames are collected and inferred. Non-key frames are estimated by interpolation to predict the detection results. This balances inference accuracy and real-time performance under the condition of limited edge computing resources, and avoids system lag caused by device overheating or insufficient computing power. Based on the deployed target detection algorithm and target tracking algorithm, real-time frame-by-frame inference is performed on each frame of the collected time-series video stream. The target detection algorithm performs forward inference on each frame of the image using a single-stage detection framework, and outputs the bounding box coordinates and category confidence of the user's hand and clothing targets, thereby transforming the original pixel information into a structured target representation that can be used for subsequent tracking and feature extraction.

[0068] Specifically, the target tracking algorithm predicts the target's position in the next frame based on the intersection-union ratio (IU) matching between the detection box in the current frame and the tracked target in the historical frames. It maintains consistent tracking of the same target across consecutive frames, thereby ensuring that the same clothing always corresponds to the same tracking identifier within the complete time window from entering the field of view to leaving the field of view, providing identity anchors for subsequent cross-frame feature aggregation.

[0069] In this embodiment, tracking the motion trajectory between the user's hand region and the clothing target region, as well as the pixel-level appearance feature sequence of the clothing target over time, specifically involves: after outputting the bounding box coordinates and category confidence scores of the user's hand and clothing target using a target detection algorithm, redundant boxes are removed through non-maximum suppression and cross-frame identity association, and a two-dimensional bounding box sequence of the user's hand region and clothing target region on the image plane is output in real time; based on this two-dimensional bounding box sequence, the relative displacement, overlap area ratio, and azimuth angle change between the hand center point and the clothing bounding box are calculated to quantify the spatial coupling degree between the hand and clothing in each frame, providing fine-grained spatial relationship features for subsequent action classification;

[0070] Specifically, a sequence of motion trajectory parameters for the hand and clothing is generated based on the pixel displacement vectors of the center points of the bounding boxes between consecutive frames. This sequence of parameters includes at least the position coordinates, displacement vectors, velocity vectors, and acceleration vectors corresponding to each frame. This is used to depict the complete dynamic change pattern of the two during the operation, enabling the temporal behavior recognition model to distinguish between two essentially opposite operational intentions: taking out and putting in, based on the motion pattern. The target tracking algorithm maintains consistent tracking of the same target across consecutive frames. For clothing targets with locked temporary tracking identifiers, pixel-by-pixel spatial sampling is performed within the pixel area covered by their bounding boxes to extract the pixel-level appearance feature sequence of the clothing target in the continuous time series.

[0071] Specifically, the appearance feature sequence includes at least the color distribution features, texture features, edge contour features, and local key point features of the clothing. The color distribution features are obtained by converting pixels within the bounding box from the RGB color space to the HSV color space and then statistically analyzing the pixel frequency distribution across each hue range, used to distinguish the main color tone and color scheme of the clothing. The texture features are obtained by encoding the pixels within the bounding box using local binary patterns and statistically analyzing their histogram distribution, used to characterize the surface texture and weave of the clothing fabric. The edge contour features are obtained by extracting edge response maps using the Canny edge detection operator and statistically analyzing the edge direction gradient histogram, used to describe the cut lines and external shape of the clothing pattern. The local key point features are obtained by detecting local key points within the bounding box area using a scale-invariant feature transform algorithm and extracting their 128-dimensional descriptors, used to capture distinctive patterns, trademarks, or specific decorative details on the clothing. These four types of features, from the four dimensions of color, texture, structure, and key points, constitute a complete description of the visual attributes of the clothing, serving together as the visual representation basis for subsequent feature vector mapping and database deduplication, ensuring sufficient feature separability between different garments.

[0072] S12: Based on the extracted temporal appearance feature sequence of clothing targets, a deep feature extraction network maps the appearance feature sequence into clothing feature vectors with preset dimensions. The clothing feature vectors are then compared one-to-one with the feature vectors of all clothing assets stored in the cloud database to obtain a similarity measurement value sequence. The current clothing target's clothing feature vector is compared one-to-one with the feature vectors of all existing clothing assets stored in the cloud database. The deep feature extraction network uses a triplet loss function to constrain features of the same clothing from different perspectives to be close to each other in the feature space, while features of different clothing are far apart. This ensures compact clustering of similar clothing in the feature space and clear boundaries between dissimilar clothing, reducing the risk of misclassifying the same clothing as different clothing due to changes in perspective or lighting differences. The similarity calculation uses a cosine similarity metric, calculating the cosine of the angle between the current feature vector and each existing feature vector as the similarity measurement value. The result is a similarity measurement value sequence containing the corresponding similarity values ​​of all existing clothing assets, thus transforming the distance relationship in the feature space into an intuitive matching score, facilitating the unified determination of subsequent thresholds.

[0073] Specifically, the maximum value in the similarity measurement sequence is compared with a preset overlap threshold. The overlap threshold is dynamically and adaptively adjusted based on the total number of existing clothing items in the database. When the database size increases, the threshold is appropriately increased to reduce the probability of false matching, and vice versa to increase the sensitivity of identifying new clothing items. If the maximum value is lower than the overlap threshold, the current clothing target is determined to be a new clothing asset appearing for the first time, triggering the subsequent new asset entry process. If the maximum value is higher than or equal to the overlap threshold, the current clothing target is determined to be a historical clothing asset already recorded in the database, and the existing clothing asset corresponding to the maximum value is taken as a candidate matching asset. At the same time, the difference between the maximum value and the second largest value is recorded as a matching confidence index. When the difference is lower than the preset confidence threshold, it indicates that there is ambiguity in the matching result. The system triggers manual review or reduces the automatic association weight of this match to avoid identity mismatch due to too many similar clothing items.

[0074] S13: Based on the similarity measurement value sequence and the preset overlap threshold, the asset feature attributes of the clothing target are obtained, the identity of the current clothing target is locked through the asset feature attributes, and a globally unique physical asset identifier is automatically generated for it in the cloud database.

[0075] If the current clothing target is determined to be a new clothing asset, then based on the extracted pixel-level appearance feature sequence, the structured attribute recognition network deployed on the device automatically identifies and extracts the structured attribute information of the clothing. The structured attribute information includes at least color category, style category, material category, and pattern category. This transforms the low-level pixel features into high-level semantic tags that are human-understandable and can be used for outfit reasoning, providing structured input fields for the subsequent outfit recommendation engine. The clothing feature vector is associated and bound with the structured attribute information, and a globally unique physical asset identifier is automatically generated for it in the cloud database, completing the seamless entry of the new clothing asset and completely freeing users from tedious manual photo taking, background removal, and tagging operations.

[0076] If the current clothing target is determined to be a historical clothing asset, the current clothing target is redirected and associated with the physical asset identifier of the candidate matching asset in the database. All historical attribute information and status information of the existing clothing are reused without repeated entry, thereby keeping the database concise and data consistent, and avoiding multiple redundant records of the same physical clothing in the system.

[0077] S14: After the identity of the current clothing target is locked, based on the relative spatial position relationship and movement trajectory between the user's hand and the clothing target obtained by tracking, the user's current operation gesture is classified by the deployed temporal behavior recognition model. According to the judgment result, the in-stock status bit corresponding to the clothing target in the database is automatically updated. Each flip of the in-stock status bit is accompanied by a timestamp record, forming a complete temporal log of the status change of clothing assets.

[0078] Specifically, based on relative spatial position relationships and motion trajectories, the relative displacement, overlapping area ratio, azimuth angle change, and hand displacement, velocity, and acceleration vectors between the hand center point and the clothing bounding box are used as input features for the temporal behavior recognition model. This allows the spatial geometric information of each frame to be aggregated into a spatiotemporal pattern reflecting the operation intention. The temporal video stream is evenly divided into multiple video segments in the time dimension. Spatial features are extracted from each segment using a two-dimensional convolutional neural network. The spatiotemporal features of multiple segments are then aggregated by a temporal fusion layer. Finally, a fully connected classification layer outputs the probability distribution of whether the operation gesture belongs to the "take out" or "put in" category. The results are further determined by referring to the trends of spatial position changes before and after the hand operation and the trends of clothing bounding box size changes. This cross-validation using multimodal contextual information improves the accuracy of gesture recognition in scenarios with blurred boundaries.

[0079] The time-series segmented confidence weighting includes:

[0080]

[0081] N: The total number of video segments into which the time-series video stream is evenly divided along the time dimension;

[0082] P k The initial probability value of the operation gesture belonging to the "take out" or "put in" category, after the spatial features of the k-th video segment are extracted by a two-dimensional convolutional neural network and output by a fully connected classification layer.

[0083] ω k The confidence weight of the k-th video segment is determined by the stability of object detection within that segment and the clarity of the hand-clothing interaction.

[0084] C k The overall confidence index within the k-th segment is obtained by weighted fusion of the mean score of the hand and clothing bounding box detection within the segment, the duration of the tracking identifier continuity, and the standard deviation of the hand-clothing overlap area. A higher value indicates that the features of the segment are more reliable.

[0085] τ: Confidence threshold, used to determine whether a segment belongs to a high confidence interval. Segments below this threshold have their weight significantly reduced in the voting process.

[0086] η: Nonlinear adjustment coefficient, which controls the steepness of the transition from low to high confidence weight;

[0087] P final The final discrimination probability after confidence-weighted voting fusion is determined as follows: when the probability is greater than 0.5, it is determined as the "into" action category; when the probability is less than 0.5, it is determined as the "out" action category. By introducing a low-confidence segment attenuation mechanism, the adverse effects of single-frame misjudgment caused by occlusion or fast movement on the final judgment result are reduced.

[0088] Specifically, based on the judgment results of the temporal behavior recognition model, the in-stock status position corresponding to the clothing target in the database is automatically updated: if the current operation gesture is determined to be a putting action, indicating that the user is hanging the clothing back or putting it in the closet, the in-stock status position will be automatically flipped to the first status value representing in-stock; if the current operation gesture is determined to be a taking action, indicating that the user is taking the clothing out of the closet to wear, the in-stock status position will be automatically flipped to the second status value representing out-of-stock; each flip of the in-stock status position is accompanied by the current system timestamp and the corresponding video frame sequence number, which are appended to the status change time sequence log of the clothing asset in chronological order to form a complete status switching history record, thereby supporting subsequent status backtracking, wearing frequency statistics and anomaly detection analysis, and providing users with a complete in-stock and out-of-stock timeline for each piece of clothing;

[0089] S15: Receive the outfit recommendation module deployed in the cloud to generate an outfit scheme. The outfit recommendation module uses the list of physical asset identifiers as the candidate recommendation SKU pool, performs outfit collaborative generation reasoning based on structured attribute information, and outputs an outfit recommendation scheme that only contains the clothing currently in stock.

[0090] Specifically, the cloud-deployed outfit recommendation module responds to outfit plan generation requests triggered by the user or automatically triggered by the system according to a preset time strategy. It sends a query command to the cloud database to check the clothing assets in stock. The query command includes at least all clothing assets whose status is defined by the first status value representing the status of being in stock. This ensures from the data source that the recommended clothing is physically present in the wardrobe. The preset time strategy includes timed triggering at a fixed time every day, delayed triggering after the user leaves the wardrobe space, and conditional triggering triggered by changes in external environmental information. This ensures that the timing of the recommendation matches the user's actual life rhythm and avoids frequently disturbing the user.

[0091] Specifically, the external environment information includes at least the weather, temperature, humidity, and UV intensity information for the day. The query command also includes filtering conditions automatically generated by the outfit recommendation module based on the external environment information and the user's historical outfit preference data, targeting one or more fields of color category, style category, and material category in the structured attribute information. This further narrows down the candidate recommendation pool based on the in-stock filtering, ensuring that the recommendation results take into account both environmental adaptability and personal style preferences. The cloud database responds to the query command by using the system timestamp when the cloud database receives the query command as a reference, and returns a list of physical asset identifiers and corresponding structured attribute information for all clothing assets marked as in-stock at the current time. This ensures the real-time consistency of the status information reading and avoids returning clothing whose status has changed after the query time.

[0092] Specifically, the outfit recommendation module uses the list of physical asset identifiers returned by the query as a candidate recommendation SKU pool. Based on a preset outfit matching rule library, it determines the feasibility of pairing or combining multiple items among the candidate SKUs. The matching rule library includes at least color matching rules, style matching rules, material matching rules, and seasonal adaptation rules. Through collaborative filtering algorithms or sequence generation models based on attention mechanisms, it sorts the feasible combinations by matching coordination scores and outputs one or more outfit recommendation schemes. This combines visual aesthetic rules with data-driven matching modes to generate recommendation results that are both in line with common sense about dressing and personalized. All clothing assets marked as out of the warehouse are automatically excluded from the candidate recommendation SKU pool. This exclusion operation is performed at the query command level rather than after the recommendation results are generated. This ensures that excluded clothing assets do not consume the inference computing resources of the outfit recommendation module and fundamentally eliminates the logical error of recommending clothing that is not actually in the wardrobe.

[0093] The matching coordination score is as follows:

[0094]

[0095] R color The color coordination score between clothing items in the candidate combination is calculated by the color matching rule library based on the hue circle angle, the relationship between contrasting and adjacent colors, and the degree of matching of brightness and saturation. The value range is [0,1].

[0096] R style The score for the coordination of the garments in the candidate combination is calculated by the pattern matching rule library based on the silhouette fit relationship, length ratio and tightness contrast of the top and bottom;

[0097] R materialThe material matching coordination score between clothing items in the candidate combination is calculated by the material matching rule library based on the degree of matching of fabric texture, thickness and surface texture.

[0098] R season The suitability score of the candidate combination with the current seasonal environment is calculated by the seasonal suitability rule library based on the degree of matching between the seasonal tag of the individual items and the season of the current date;

[0099] ω c ω s ω m ω season These are the adaptive weighting coefficients corresponding to the four dimensions of scoring mentioned above. Each weighting coefficient is dynamically adjusted based on the user's historical outfit feedback data and satisfies ω. c +ω s +ω m +ω season =1;

[0100] This technical solution continuously observes the user's natural picking and placing behavior using a single static camera. Without the user noticing, it utilizes edge-side target detection and tracking algorithms to lock onto the interaction between the hand and clothing in real time. It automatically distinguishes between new and historical clothing using appearance feature extraction and database similarity comparison, and accurately determines the two types of gestures—taking out and putting in—based on an action recognition model. This drives the state machine to complete the real-time flipping of the inventory status, eliminating the system's reliance on user-initiated operations and integrating asset management into the user's daily life habits. Simultaneously, the system uses a unique physical asset identifier for each garment as its core, automatically maintaining a time-series log of state changes with timestamps. This provides a structured foundation for subsequent data value-added services such as wear frequency analysis and styling effect tracking. This not only lowers the deployment threshold and usage cost of smart wardrobes but also lays the foundation for personalized outfit recommendations based on real inventory data, increasing user trust and satisfaction with smart outfit services.

[0101] Reference Figure 3 As shown, the apparel asset seamless management system based on time-series video streams is used to execute the aforementioned apparel asset seamless management method. The management system includes:

[0102] The video acquisition module is used to continuously acquire the time-series video stream during the user's operation of taking and putting things in the wardrobe and transmit it to the terminal.

[0103] The edge processing module is used to perform real-time target detection and tracking on the time-series video stream, extract the pixel-level appearance feature sequence of clothing targets and map it into clothing feature vectors, identify the structured attribute information of clothing, and determine the category of taking out or putting in based on the user's current operation gesture.

[0104] The cloud database module is used to store the physical asset identifier, feature vector, structured attribute information and inventory status bits of all clothing assets, and maintain the status change time sequence log for each garment.

[0105] The deduplication and association module is used to compare the feature vector of clothing with the existing feature vector in the database to determine whether the current clothing is a new asset or a historical asset, and trigger the entry of new assets or association with existing identifiers accordingly.

[0106] The status management module is used to flip the in-stock status bit of the corresponding clothing to in-stock or out-of-stock status based on the operation gesture judgment result.

[0107] The outfit recommendation module responds to outfit generation requests by querying all in-stock clothing assets from the database and using the query results as a candidate recommendation pool to generate outfit recommendation schemes. It continuously observes the user's natural picking and placing behavior through a single static camera. Under the premise that the user is completely unaware, it uses edge-side object detection and tracking algorithms to lock the interaction between the hand and the clothing in real time. It automatically distinguishes between new and old clothing by extracting appearance features and comparing similarity with the database, and accurately determines the two types of operation gestures of taking out and putting in based on the action recognition model. It drives the state machine to complete the real-time flipping of the in-stock status position, eliminating the system's dependence on the user's active operation and integrating the asset management process into the user's daily life habits.

[0108] Specifically, the edge processing module includes:

[0109] The target detection and tracking unit is used to perform target detection in each frame of the image, maintain consistent tracking of the same clothing target across consecutive frames, and output the motion trajectory of the hand and clothing.

[0110] The feature extraction unit is used to extract the pixel-level appearance feature sequence of the clothing target in time, map the appearance feature sequence into clothing feature vectors, and identify the structured attribute information of the clothing.

[0111] The behavior recognition unit is used to determine the category of user operation gestures for taking out or putting in based on the motion trajectory.

[0112] The plagiarism detection and association module includes:

[0113] The similarity calculation unit is used to calculate the similarity measure between the clothing feature vector and each existing feature vector in the database;

[0114] The threshold comparison unit is used to compare the maximum value in the similarity measurement value sequence with the preset overlap threshold and output a judgment signal for new clothing assets or historical clothing assets.

[0115] The asset entry unit is used to generate a unique physical asset identifier and create a new record in the database in response to the new clothing asset determination signal;

[0116] The asset association unit is used to associate the current clothing target with an existing physical asset identifier in the database in response to the historical clothing asset determination signal.

[0117] The status management module includes:

[0118] The state flipping unit is used to send a state update instruction to the database based on the operation gesture determination result, and update the database status bit to "in the database" or "out of the database".

[0119] The log recording unit is used to record the state switching information and timestamp each time the state is updated, and write it into the state change log of the corresponding garment.

[0120] The outfit recommendation module includes:

[0121] The query building unit is used to generate query instructions containing library status filtering conditions in response to the outfit generation request;

[0122] The recommendation generation unit uses the clothing list returned by the query as a candidate recommendation pool, judges the feasibility of matching according to the matching rules, and outputs matching recommendations based on the coordination score. Through the collaborative work of a lightweight convolutional neural network and a segmented temporal behavior recognition model, it can maintain stable real-time inference performance under the condition of limited computing resources. Feature vector retrieval and state filtering of cloud database ensure that the matching recommendation engine generates matching based only on the clothing assets currently in the database, and the recommendation results are always consistent with physical reality.

[0123] In this embodiment, the entire operation process can be automated by computer control. In each operation stage, sensors can be set up to provide signal feedback and ensure that the steps are performed sequentially. These are all conventional knowledge of current automation control, and will not be elaborated on in this embodiment.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for seamless management of apparel assets based on time-series video streams, applied to seamless asset management devices, specifically includes the following steps: S11: A single static camera deployed in the wardrobe captures a time-series video stream of the user's actions of taking and putting clothes in the wardrobe. The system performs real-time frame-by-frame reasoning on each frame of the captured time-series video stream, tracks the motion trajectory between the user's hand area and the target area of ​​the clothing, and the pixel-level appearance feature sequence of the clothing target in the time series. Based on the motion trajectory, the system initially determines the behavior category of the clothing target involved in the current operation in the time series. S12: Based on the extracted appearance feature sequence of the clothing target in time, the appearance feature sequence is mapped into a clothing feature vector with a preset dimension through a deep feature extraction network. The clothing feature vector is then compared with the feature vectors of all clothing assets stored in the cloud database to obtain a similarity measurement value sequence. S13: Based on the similarity measurement value sequence and the preset overlap threshold, the asset feature attributes of the clothing target are obtained, the identity of the current clothing target is locked through the asset feature attributes, and a globally unique physical asset identifier is automatically generated for it in the cloud database. S14: After the identity of the current clothing target is locked, based on the relative spatial position relationship and movement trajectory between the user's hand and the clothing target obtained by tracking, the user's current operation gesture is classified by the deployed temporal behavior recognition model. According to the judgment result, the in-stock status bit corresponding to the clothing target in the database is automatically updated. Each flip of the in-stock status bit is accompanied by a timestamp record, forming a complete temporal log of the status change of the clothing asset. S15: Receive an outfit recommendation request initiated by the outfit recommendation module deployed in the cloud. The outfit recommendation module uses the list of physical asset identifiers as the candidate recommendation SKU pool, performs outfit collaborative generation reasoning based on structured attribute information, and outputs an outfit recommendation scheme that only contains clothing currently in stock.

2. The method for seamless management of apparel assets based on time-series video streams as described in claim 1, characterized in that, In S11, a single static camera deployed inside the wardrobe captures a time-series video stream of the user's actions of taking and placing items in the wardrobe. Real-time frame-by-frame inference is performed on each frame of the captured time-series video stream, including: A single static camera is fixedly deployed in the top corner of the wardrobe or in an external position facing the wardrobe door opening. Its optical axis is directed towards the clothing retrieval and placement area of ​​the wardrobe. The camera continuously captures a time-series video stream of the complete operation process of the user opening the wardrobe door, taking out and placing clothes, and closing the wardrobe door at a preset frame rate of 15 to 30 frames per second. The preset frame rate is dynamically and adaptively adjusted according to the current computing load of the end device. When the load is lower than the first threshold, the first frame rate value is used. When the load is higher than the first threshold but lower than the second threshold, the second frame rate value is used. When the load is higher than the second threshold, a frame skipping strategy is triggered, and only key frames are captured and inferred. Non-key frames are estimated by interpolation. Based on the deployed target detection and target tracking algorithms, real-time frame-by-frame inference is performed on each frame of the acquired time-series video stream. The target detection algorithm performs forward inference on each frame of the image using a single-stage detection framework, and outputs the bounding box coordinates and category confidence of the user's hand and clothing targets. The target tracking algorithm predicts the target's position in the next frame based on the intersection-union ratio (IU) matching between the detection box in the current frame and the tracked target in the historical frames, thus maintaining consistent tracking of the same target across consecutive frames.

3. The method for seamless management of apparel assets based on time-series video streams as described in claim 2, characterized in that, Tracking the motion trajectory between the user's hand area and the target area of ​​the clothing, as well as the pixel-level appearance feature sequence of the clothing target over time, including: After the target detection algorithm outputs the bounding box coordinates and category confidence scores of the user's hand and clothing targets, redundant boxes are removed by non-maximum suppression and cross-frame identity association. The two-dimensional bounding box sequence of the user's hand region and clothing target region on the image plane is output in real time. Based on the two-dimensional bounding box sequence, the relative displacement, overlap area ratio and azimuth change between the hand center point and the clothing bounding box are calculated. The motion trajectory parameter sequence of the hand and clothing is generated based on the pixel displacement vector of the bounding box center point between consecutive frames. The target tracking algorithm maintains consistent tracking of the same target across frames. For clothing targets with locked temporary tracking identifiers, pixel-by-pixel spatial sampling is performed within the pixel area covered by their bounding box to extract the pixel-level appearance feature sequence of the clothing target in the continuous time series. The appearance feature sequence includes at least the color distribution features, texture features, edge contour features, and local key point features of the clothing. The color distribution features are obtained by converting the pixels within the bounding box from the RGB color space to the HSV color space and then calculating the pixel frequency distribution of each hue range. The texture features are obtained by performing local binary pattern encoding on the pixels within the bounding box and calculating their histogram distribution. The edge contour features are obtained by extracting the edge response map using the Canny edge detection operator and calculating the edge direction gradient histogram. The local key point features are obtained by detecting local key points within the bounding box region and extracting their 128-dimensional descriptors using the scale-invariant feature transform algorithm.

4. The method for seamless management of apparel assets based on time-series video streams as described in claim 3, characterized in that, In S12, the similarity between the clothing feature vector and the feature vectors of all clothing assets stored in the cloud database is calculated one by one to obtain a sequence of similarity metrics, including: The similarity of the current clothing feature vector with the feature vectors of all existing clothing assets stored in the cloud database is calculated one by one. The triple loss function is used to constrain the features of the same clothing from different perspectives to be close to each other in the feature space and the features of different clothing to be far apart. The similarity calculation adopts the cosine similarity measurement method, and the cosine value of the angle between the current feature vector and each existing feature vector is calculated as the similarity measurement value. After calculation, a similarity measurement value sequence containing the corresponding similarity values ​​of all existing clothing assets is obtained. The maximum value in the similarity metric sequence is compared with a preset overlap threshold, which is dynamically and adaptively adjusted according to the total number of existing clothes in the database. If the maximum value is lower than the overlap threshold, the current clothing target is determined to be a new clothing asset appearing for the first time, triggering the subsequent new asset entry process; if the maximum value is higher than or equal to the overlap threshold, the current clothing target is determined to be a historical clothing asset that has been recorded in the database, and the existing clothing asset corresponding to the maximum value is taken as a candidate matching asset, while the difference between the maximum value and the second largest value is recorded as a matching confidence index.

5. The method for seamless management of apparel assets based on time-series video streams as described in claim 4, characterized in that, In S13, the identity of the current clothing target is locked through asset characteristic attributes, and a globally unique physical asset identifier is automatically generated for it in the cloud database, including: If the current clothing target is determined to be a new clothing asset, then based on the extracted pixel-level appearance feature sequence, the structured attribute information of the clothing is automatically identified and extracted through the structured attribute recognition network deployed on the edge. The clothing feature vector is associated and bound with the structured attribute information, and a globally unique physical asset identifier is automatically generated for it in the cloud database, thus completing the seamless entry of the new clothing asset. If the current clothing target is determined to be a historical clothing asset, then the current clothing target is redirected and associated with the existing physical asset identifier in the database through the physical asset identifier of the candidate matching asset, and all historical attribute information and status information of the existing clothing are reused.

6. The method for seamless management of apparel assets based on time-series video streams as described in claim 5, characterized in that, In S14, based on the tracked relative spatial relationship and movement trajectory between the user's hand and the clothing target, a deployed temporal behavior recognition model is used to classify the user's current gesture, including: Based on the relative spatial position relationship and motion trajectory, the relative displacement, overlapping area ratio, azimuth angle change, and hand displacement vector, velocity vector, and acceleration vector between the hand center point and the clothing bounding box are used as input features of the temporal behavior recognition model. The temporal video stream is uniformly divided into multiple video segments in the time dimension. Spatial features are extracted for each segment through a two-dimensional convolutional neural network. Then, the spatiotemporal features of multiple segments are aggregated through a temporal fusion layer. Finally, the probability distribution of the operation gesture belonging to the take-out or put-in category is output through a fully connected classification layer. The result is determined with the help of the trend of spatial position change before and after the hand operation and the trend of clothing bounding box size change. Based on the determination result of the time-series behavior recognition model, the in-stock status position corresponding to the clothing target in the database is automatically updated. If it is determined that the current operation gesture belongs to the putting action category, indicating that the user is hanging the clothing back or putting it into the closet, the in-stock status position is automatically flipped to the first status value representing in-stock. If the current gesture is determined to be a take-out action, indicating that the user is taking clothes out of the closet to wear, the in-stock status will be automatically flipped to the second status value representing out-of-stock. Each flip of the in-stock status bit is accompanied by the current system timestamp and the corresponding video frame number record, which are appended to the status change time sequence log of the clothing asset in chronological order to form a status switching history record.

7. The method for seamless management of apparel assets based on time-series video streams as described in claim 6, characterized in that, In S15, an outfit recommendation module deployed in the cloud is received, which generates an outfit scheme. This module uses a list of physical asset identifiers as its candidate recommendation SKU pool, including: The cloud-deployed outfit recommendation module responds to outfit scheme generation requests triggered by users or automatically triggered by the system according to a preset time strategy, and sends an instruction to the cloud database to query clothing assets in stock. The query instruction includes at least all clothing assets whose filtering condition is that the stock status bit value is a first status value representing the stock status. The preset time strategy includes timed triggering at a fixed time every day, delayed triggering after the user leaves the wardrobe space, and conditional triggering triggered by changes in external environmental information. The external environment information includes at least the weather information, temperature information, humidity information, and ultraviolet intensity information for the day. The query instruction also includes filtering conditions for one or more fields of color category, style category, and material category in the structured attribute information, which are automatically generated by the outfit recommendation module based on the external environment information and the user's historical outfit preference data.

8. The method for seamless management of apparel assets based on time-series video streams as described in claim 7, characterized in that, Send a query command to the cloud database to check the inventory of clothing assets, including: In response to a query command, the cloud database returns a list of physical asset identifiers and corresponding structured attribute information for all clothing assets marked as in stock at the current moment, based on the system timestamp when the cloud database receives the query command. The outfit recommendation module uses the list of physical asset identifiers returned by the query as a candidate recommendation SKU pool. Based on the preset outfit matching rule library, it determines the feasibility of pairing or multi-item combinations of candidate SKUs. The matching rule library includes at least color matching rules, style matching rules, material matching rules and season matching rules. It also outputs one or more outfit recommendation schemes by sorting the feasible combinations according to the matching coordination score through collaborative filtering algorithm or sequence generation model based on attention mechanism. All clothing assets marked as out of stock are automatically excluded from the candidate recommended SKU pool to ensure that the recommendation results are consistent with the actual physical asset status in stock in real time.

9. A seamless apparel asset management system based on time-series video streams, characterized in that: For implementing the apparel asset seamless management method according to any one of claims 1-8, the management system includes: The video capture module is used to continuously capture the time-series video stream of the user's actions of taking and putting things in the wardrobe and transmit it to the terminal. The edge processing module is used to perform real-time target detection and tracking on the time-series video stream, extract the pixel-level appearance feature sequence of clothing targets and map it into clothing feature vectors, identify the structured attribute information of clothing, and determine the category of taking out or putting in based on the user's current operation gesture. The cloud database module is used to store the physical asset identifier, feature vector, structured attribute information and inventory status bits of all clothing assets, and maintain the status change time sequence log for each garment. The deduplication and association module is used to compare the feature vector of the clothing with the existing feature vector in the database, determine whether the current clothing is a new asset or a historical asset, and trigger the entry of new assets or association with existing identifiers accordingly. The status management module is used to flip the in-stock status bit of the corresponding clothing to in-stock or out-of-stock status based on the operation gesture judgment result. The outfit recommendation module is used to respond to outfit generation requests by querying all clothing assets in the database and using the query results as a candidate recommendation pool to generate outfit recommendation schemes.

10. The seamless apparel asset management system based on time-series video streams as described in claim 9, characterized in that, The end-side processing module includes: The target detection and tracking unit is used to perform target detection in each frame of the image, maintain consistent tracking of the same clothing target across consecutive frames, and output the motion trajectory of the hand and clothing. The feature extraction unit is used to extract the pixel-level appearance feature sequence of the clothing target in time, map the appearance feature sequence into clothing feature vectors, and identify the structured attribute information of the clothing. The behavior recognition unit is used to determine the category of taking out or putting in the user's operation gesture based on the motion trajectory.

Images